Predicted Interaction Energy Between Organic Molecule And Water Or Organic Molecule And Ion By Deep Tensor Neural Network

Fast and accurate calculation of molecular interaction energy plays an important role in efficient and accurate molecular dynamics simulation.In the existing methods,the interaction energy based on the molecular force field can be calculated quickly,but the accuracy is not high.The interaction energy based on quantum chemistry has high precision,but it has higher requirements on computing resources and computing time.Therefore,it is necessary to develop a fast and high precision interaction energy computing framework.In recent years,Deep Learning has obtained breakthrough in a diversity of disciplines and application fields.In this paper,we propose a scheme called Deep Tensor Neural Network(DTNN)to predict the interaction energy which is between organic molecule and water or organic molecule and ion and to predict its physicochemical properties with relative high-accuracy.Deep neural network constructs multi hidden-layer network by convoluting and pooling and so on,it trains node weights in the hidden layer through a large amount of data and realizes the learning of hidden characteristics in the data.DTNN is the further development of deep neural network.In the DTNN,the input is converted into a set of tensor by basis function.DTNN extracts valid information by convolution.First,we generate a great deal of structure files of organic molecule and water or organic molecule and ion within the fixed range.And we obtain the interaction energy between two molecules system with B3 LYP method and 6-31+G* basis set by Quantum Chemical computing software and generate the data set.Then,the network inputs of DTNN include pair-wise inter-distance and nuclear charges of atoms of each molecule,and the network outputs include interaction energy of two molecules.DTNN is trained,valided,and tested by tensorflow and predict the interaction energy of the systems between organic molecule and water or organic molecule and ion on the GPU.In the paper,we predict the interaction energy of conformation in a great deal of systems by the DTNN and compare the corresponding data to QM.The prediction results demonstrate good correlations.The average value of the absolute prediction error and the standard deviation of the absolute prediction error are within the scope of acceptable,and the overall prediction accuracy is very high.Compared to QM calculation,the scheme we proposed reduces the computational time considerably with acceptable precision loss.